hERG: the silent killer of kinase programs

More clinical candidates die from hERG than from any other off-target liability. The ones that survive into Phase III with a hERG signal carry a black-box label for the rest of the franchise. The cardiac risk gets undersold in early discovery because it’s invisible to a simple potency assay — but the structural pattern that causes it is so common that medicinal chemists eventually learn to spot it from the SMILES alone.

What hERG actually is

hERG (the human Ether-à-go-go Related Gene, IKr channel) is a voltage-gated potassium channel that handles the rapid component of cardiac repolarization. When you block it, the action potential gets longer. On an ECG that shows up as QT prolongation. QT prolongation can trigger torsades de pointes — a polymorphic ventricular tachyarrhythmia that occasionally degenerates into ventricular fibrillation and sudden cardiac death.

The history of the field is littered with hERG-driven withdrawals: terfenadine (Seldane, 1997), cisapride (Propulsid, 2000), grepafloxacin (Raxar, 1999). Each one taught the FDA to take cardiac liability more seriously. By the time vandetanib was approved in 2011, the agency was requiring thorough QT studies for every new molecular entity in oncology. That requirement hasn’t loosened.

The pharmacophore that gets you in trouble

Aronov’s 2005 review crystallized the structural pattern. A high hERG-block probability needs three ingredients:

• A basic nitrogen, usually in a piperidine, piperazine, or tertiary amine. Protonated under physiological pH, anchors the molecule into the central cavity by hydrogen-bonding to Tyr652 and stacking against Phe656.
• Two or more lipophilic aromatic rings, separated by a few rotatable bonds. The rings make π-stacking interactions deep in the channel pore.
• Overall logP > 3.5. Lipophilic enough to find the hydrophobic pore in the first place.

Look at terfenadine, cisapride, astemizole, sertindole. They all match. Look at most marketed kinase inhibitors. They mostly match too — kinase inhibitors love basic amines for solubility, love aromatic rings for the ATP-pocket hinge, and tend to settle around logP 3-5. That’s why hERG screening is mandatory for every kinase program.

What the prediction methods actually do

Three layers of evidence, ordered by cost:

• Rule-based heuristics (SMARTS pattern matching for the Aronov pharmacophore). Free, instant, decent recall, terrible precision. Will flag almost every kinase inhibitor as “hERG high” — including ones that don’t actually block hERG — because the pattern is so common.
• ML predictors (admet-ai, ADMETLab, Schrödinger’s QSAR). Trained on patch-clamp data from Karim et al. and the Therapeutics Data Commons. Output is a continuous probability, 0-1. Reasonable accuracy (~80% AUC), with the well-known caveat that compounds outside the training distribution can fail in either direction.
• Wet-lab patch clamp on hERG-expressing HEK293 cells. The gold standard. ~$5K per compound at a CRO, two-week turnaround. Required for any compound advancing past lead-opt.

Try the prediction yourself

Liganx’s ADMET panel runs the admet-ai Chemprop ensemble on every compound after a successful dock. It returns a continuous hERG probability that maps to low/medium/high tiers (0.5 cutoff is the literature standard). The evidence string shows the raw probability so you can see how close to the cutoff a borderline compound was.

Open Studio and dock any candidate. After the run completes, click the violet ⚕ ADMET pill on a result row to see the risk profile. Three colored dots summarize hERG, DILI, and CYP3A4 risk at a glance — emerald for low, amber for medium, rose for high. If your candidate is rose-rose-rose, it’s not necessarily dead, but it’s a strong signal to redesign before sinking the synthesis cost.

Liganx brings molecular docking online into the browser and runs the ADMET panel on every pose. Using molecular docking and the cardiac-risk readout together is how you catch a hERG liability before the synthesis cost.

Primary sources

• Aronov AM. Predictive in silico modeling for hERG channel blockers. Drug Discov Today 10, 149-155 (2005). doi:10.1016/S1359-6446(04)03278-7
• Karim A, Lee M, Balle T, Sattar A. CardioTox net: a robust predictor for hERG channel blockade based on deep learning meta feature ensembles. J Cheminform 13, 60 (2021). doi:10.1186/s13321-021-00541-z
• Swain CG, Lewis ML. Toxic effects of pharmaceuticals on the hERG channel: a regulatory perspective. Br J Pharmacol 159, 5-12 (2010).